In today’s era of data-driven healthcare, the benefits of machine learning (ML) are only beginning to be utilised. However, with the growing power of ML models, there has been a parallel concern about transparency. Especially in critical decisions like cancer treatments, it’s vital to understand how decisions are made, shedding light on potential biases.

Oesophageal cancer (OC) is complex, with treatment options often depending on a myriad of factors, including patient age. Though age is a central determinant in this decision-making process, the precise way in which it influences decisions is not well characterised. This obscurity potentially leads to health inequalities among patients. Our research utilised the power of interpretable ML to clarify how age, along with other clinical factors, determines treatment choices.

Thus far we have looked retrospectively at 893 patients from 2010-2022 at our unit. Using the random forests (RF) classifier, we predicted treatment routes, such as neoadjuvant chemotherapy (NACT) followed by surgery, neoadjuvant chemoradiotherapy (NACRT) followed by surgery, surgery alone, and palliative management.

Understanding how much each variable, especially age, influenced these decisions was paramount. For this, we adopted Variable Importance and Partial Dependence analyses with the help of Assoc. Prof Arya Farahi from UT, Austin, solidifying our collaborative efforts in ensuring methodological rigor.

We found considerable influence of age on the RF model (17.2% of total importance). For patients >75, there was a noticeable shift in predicted treatment probabilities, impacting the treatment modalities. Patients between 75-85 years had increased probabilities for surgery-alone and palliative therapies. Conversely, they had a reduced likelihood for NACT/NACRT. Additionally, other factors, like staging characteristics and performance status, also played pivotal roles depending on the patient’s age.

Recognitions & Next Steps

 

Our efforts and results were recognized at the Roux Group 2023, where our PhD student, Nav Thavanesan (pictured below on the right), presented the work and deservedly bagged the 1st Prize. Our findings have also been presented at the UKIOG, ISDE and BASO. Our preliminary findings have led to a published manuscript in the ESJO (European Journal of Surgical Oncology).

For our next steps, we’re engaging both patients and clinicians in focus groups and questionnaires to assess AI acceptability of a decision making tool. We will also continue a series of 3 RRI (responsible research and innovation) meetings involving patients.

Our research is more than just another study; it’s about ensuring fairness in oesophageal treatment decisions to lead to heath equality. By using interpretable ML techniques, we’re not only allowing for streamlined workflows within cancer care but also ensuring that advancements are understood and transparent by all stakeholders.

This is the TAS funded project.